The difference is so glaring it’s as if it got silently switched to image-1. I’ve been working on a storyboard builder app and creating hundreds of test images over the last few weeks using a catalog of very detailed prompts which has produced consistent and high quality results. As of the last 24 hours it’s delivering dramatically worse results. Poor prompt adherence, artifacts/extra limbs, uncanny faces that look like the AI slop faces from a year ago. Something strange has happened.
Welcome to the community @shmitterwink !
That is something I’ve also noticed since yesterday. I usually have quality:low, but since yesterday I have needed to switch to quality: medium and the image quality got better.
Do you experience that quality in images are worse, no matter which image quality you have chosen?
Thank you. I’m happy to see I’m not going crazy.
It’s especially true for low though I do notice the difference across the board. I was getting perfectly fine results on low and now it’s behaving like 2024 Grok Imagine.
I’d probably collect a few more examples first (something I will do and keep an extra eye on the quality in images). It would be useful to compare the same prompt with the same settings across low and medium and see whether more users can reproduce it before assuming it’s a permanent model regression. Personally I think, this is a temporarily glitch.
But if you want to, please keep us updated if the issue remains or if GPT-image-2 is generating the quality it suppose to.
I’m having the same issue since yesterday. Getting image quality of two years ago.
When GPT Image 2 was released I switched all of my production traffic from Nano Banana Pro immediately as the improvement was obvious across all my test cases. Since approximately 1 month ago the quality has clearly degraded even at the highest quality setting (high). If someone at OpenAI is interested in investigating I’d be happy to share more information and some examples. By now, I guess I’m switching back to Nano Banana Pro.
I’ve concluded the severe reduction in quality is an intentional move on OpenAI’s part and is not going to be addressed.
It’s intentional alright, and all frontier companies are doing this to save compute. You’re paying full price and getting served up quantized and degraded models. Image-2 is nothing like it was when it first dropped. Every gen I get is garbled, low quality, burned and artifacted now across API and the site.
Has it improved for you yet?